Tianshu Harness
A terminal-and-desktop coding-agent runtime built around a cognitive virtual machine that externalizes goals, state, evidence and permissions from the conversation history, so a model's own claim of being finished must pass runtime-checked delivery evidence.
Tianshu Harness is a TypeScript coding-agent runtime whose terminal TUI and desktop GUI share one kernel, aimed at a model delivering stably: not drifting from the goal, not claiming completion it cannot show evidence for, and not needing repeated re-verification of the same claim. It does this with a Cognitive Virtual Machine, a layer between the model and real actions with 72 runtime hooks across five phases, keeping goals, state, evidence, resources, permissions and termination conditions external to the conversation.
A task contract pins the overall goal so that a delivery gate requires runtime evidence, tests, diffs, verification commands, before work is considered complete, and a convergence check independently judges whether the agent's reasoning is still making progress. It ships a purpose-built ANSI terminal UI and a Tauri desktop app for macOS, Windows and Linux sharing the same agent kernel, and includes a prefix-caching engine the project reports reaching a 98 to 99 percent steady-state cache hit rate on long sessions, with additional tuning for DeepSeek V4.
It supports DeepSeek, GLM, Claude, Codex, Grok, MiniMax and MiMo as underlying models; its original development codename was Rivet, and the CLI's data directory remains ~/.rivet.